1 hour ago
22
Meituan, the company best known for delivering dumplings to your door, just dropped one of the most significant AI models to come out of China. LongCat-2.0, released on June 30, packs 1.6 trillion parameters and was trained entirely on a cluster of 50,000 domestically manufactured GPU chips. No NVIDIA hardware in sight.
The US has spent the last few years tightening export controls on advanced AI chips to China, operating under the assumption that cutting off access to top-tier NVIDIA silicon would slow Beijing’s AI ambitions.
What Meituan actually built
LongCat-2.0 uses a mixture of experts (MoE) architecture, which is a design approach where the model has 1.6 trillion total parameters but only activates roughly 48 billion of them for any given task. The model was pretrained on over 30 trillion tokens and supports a context window of 1 million tokens. On the Terminal Bench 2 coding benchmark, LongCat-2.0 scored 70.8, placing it in competitive territory with leading models globally.
The 50,000-chip domestic cluster represents what appears to be the largest reported training run using non-NVIDIA hardware.
Moonshot AI’s parallel push
Moonshot AI has been iterating on its Kimi model series, with variants reaching 1 trillion parameters trained on 15.5 trillion tokens. The Kimi K2 series features around 32 billion active parameters and incorporates a proprietary technique called the MuonClip optimizer, which helps maintain stability during training.
While the Kimi models and Meituan’s LongCat-2.0 are separate efforts from different companies, the research notes there is no direct association between Moonshot AI’s efforts and the large domestic training cluster used by Meituan.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·